---
title: "Giant product street stunt — Meleon Skill"
description: "Make an 8 to 10 second vertical fake out-of-home (FOOH) stunt of your own product, giant, in a city street, filmed like a phone clip. Pick a stunt (out of a billboard, rolling down a hill street, parked on a rooftop, wrapping a tram, rising from a harbour) and a type of street. Plans two 5 second shots (a wide establishing shot, then a low-angle hero moment), locks the product's real shape with keyframes made from your photo, tests the move with cheap 480p drafts, renders the finals with Kling 3 pro and native street sound, and joins them with ffmpeg into one 9:16 MP4 with a small \"AI-made\" tag drawn locally and on by default. Comes with a caption that says it is CGI. No real brands, billboards or landmarks, nothing damaged, no panic, tiny pedestrians with unreadable faces, never passed off as real. For brands, product marketers and creators."
url: https://meleon.ai/skills/giant-product-street-stunt/
author: "Meleon AI"
---

# Giant product street stunt

> Put your product into a city street as a giant, phone-filmed stunt clip of 8 to 10 seconds with street sound and an AI-made tag.

- Type: Video Skill
- Category: Video & animation
- Version: 1.0.0 (updated 2026-11-13)
- Made by: Meleon AI (official)
- Price: Free to download. Outside costs: About $2.70 for a stunt on fal (keyframes $0.30 on Nano Banana Pro edit, 480p drafts $0.72 on Seedance 2.0 mini, two 5 s Kling 3 pro finals with sound $1.68), about $4.95 for the sample run with re-rolls (drafts skipped); Seedance 2.5 finals instead make it about $5.75. Billed to your own fal account; quote the total and get a yes before the first paid call.
- Runs in: Claude Code, Codex and Cursor
- Page: https://meleon.ai/skills/giant-product-street-stunt/
- Download: https://meleon.ai/downloads/giant-product-street-stunt.zip
- Manifest: https://meleon.ai/skills/giant-product-street-stunt.json

## What it does

Make an 8 to 10 second vertical fake out-of-home (FOOH) stunt of your own product, giant, in a city street, filmed like a phone clip. Pick a stunt (out of a billboard, rolling down a hill street, parked on a rooftop, wrapping a tram, rising from a harbour) and a type of street. Plans two 5 second shots (a wide establishing shot, then a low-angle hero moment), locks the product's real shape with keyframes made from your photo, tests the move with cheap 480p drafts, renders the finals with Kling 3 pro and native street sound, and joins them with ffmpeg into one 9:16 MP4 with a small "AI-made" tag drawn locally and on by default. Comes with a caption that says it is CGI. No real brands, billboards or landmarks, nothing damaged, no panic, tiny pedestrians with unreadable faces, never passed off as real. For brands, product marketers and creators.

## Made for

- Brands
- Product marketers
- Creators
- Social media teams
- Agencies

## Before you start

- You give it: Product photo + a stunt and a street
- You need: Node.js 18 or newer and ffmpeg with ffprobe for the scripts; sharp for the AI-made tag; your own FAL_KEY only if you run scripts/fal-run.mjs.
- Cost per run: About $2.70 for a stunt on fal (keyframes $0.30 on Nano Banana Pro edit, 480p drafts $0.72 on Seedance 2.0 mini, two 5 s Kling 3 pro finals with sound $1.68), about $4.95 for the sample run with re-rolls (drafts skipped); Seedance 2.5 finals instead make it about $5.75. Billed to your own fal account; quote the total and get a yes before the first paid call.
- Makes: 2 keyframes + 2 Kling 3 pro clips with sound, joined into one 9:16 MP4 of 8 to 10 s, 720p, with a local AI-made tag

## Install

Run the command for your agent in a terminal. It downloads https://meleon.ai/downloads/giant-product-street-stunt.zip and unzips it into the agent's personal Skills folder.

### Claude Code

```sh
mkdir -p ~/.claude/skills && curl -fsSL https://meleon.ai/downloads/giant-product-street-stunt.zip -o /tmp/giant-product-street-stunt.zip && unzip -o /tmp/giant-product-street-stunt.zip -d ~/.claude/skills
```

Then ask Claude Code to use it: `/giant-product-street-stunt`.

### Codex

```sh
mkdir -p ~/.agents/skills && curl -fsSL https://meleon.ai/downloads/giant-product-street-stunt.zip -o /tmp/giant-product-street-stunt.zip && unzip -o /tmp/giant-product-street-stunt.zip -d ~/.agents/skills
```

Then ask Codex to use it: `$giant-product-street-stunt`.

### Cursor

```sh
mkdir -p ~/.cursor/skills && curl -fsSL https://meleon.ai/downloads/giant-product-street-stunt.zip -o /tmp/giant-product-street-stunt.zip && unzip -o /tmp/giant-product-street-stunt.zip -d ~/.cursor/skills
```

Then ask Cursor to use it: `/giant-product-street-stunt`.

## Example request

```text
/giant-product-street-stunt our trainer, giant, leaning out of a billboard on a city avenue, 9:16
```

In Codex, start with `$giant-product-street-stunt` instead of `/giant-product-street-stunt`.

## Files

Every file in the package, readable before you install:

- [SKILL.md](https://meleon.ai/skills/giant-product-street-stunt/files/SKILL.md) (14.2 KB)
- [references/output-spec.md](https://meleon.ai/skills/giant-product-street-stunt/files/references/output-spec.md) (6.0 KB)
- [references/prompt-patterns.md](https://meleon.ai/skills/giant-product-street-stunt/files/references/prompt-patterns.md) (8.3 KB)
- [references/safety-and-honesty.md](https://meleon.ai/skills/giant-product-street-stunt/files/references/safety-and-honesty.md) (3.4 KB)
- [references/shot-grammar.md](https://meleon.ai/skills/giant-product-street-stunt/files/references/shot-grammar.md) (3.9 KB)
- [references/stunt-menu.md](https://meleon.ai/skills/giant-product-street-stunt/files/references/stunt-menu.md) (3.8 KB)
- [assets/caption-template.md](https://meleon.ai/skills/giant-product-street-stunt/files/assets/caption-template.md) (1.5 KB)
- [assets/example-input.jpg](https://meleon.ai/skills/giant-product-street-stunt/files/assets/example-input.jpg/) (155.0 KB)
- [assets/generation-input.json](https://meleon.ai/skills/giant-product-street-stunt/files/assets/generation-input.json) (3.9 KB)
- [assets/poster.webp](https://meleon.ai/skills/giant-product-street-stunt/files/assets/poster.webp/) (69.5 KB)
- [assets/preview.mp4](https://meleon.ai/skills/giant-product-street-stunt/files/assets/preview.mp4) (6.6 MB)
- [assets/prompt-template.md](https://meleon.ai/skills/giant-product-street-stunt/files/assets/prompt-template.md) (6.2 KB)
- [assets/stunts.json](https://meleon.ai/skills/giant-product-street-stunt/files/assets/stunts.json) (14.2 KB)
- [scripts/assemble-stunt.mjs](https://meleon.ai/skills/giant-product-street-stunt/files/scripts/assemble-stunt.mjs) (13.7 KB)
- [scripts/build-plan.mjs](https://meleon.ai/skills/giant-product-street-stunt/files/scripts/build-plan.mjs) (6.7 KB)
- [scripts/build-requests.mjs](https://meleon.ai/skills/giant-product-street-stunt/files/scripts/build-requests.mjs) (12.8 KB)
- [scripts/check-stunt.mjs](https://meleon.ai/skills/giant-product-street-stunt/files/scripts/check-stunt.mjs) (14.2 KB)
- [scripts/fal-run.mjs](https://meleon.ai/skills/giant-product-street-stunt/files/scripts/fal-run.mjs) (8.4 KB)

## Versions

- 1.0.0 (2026-11-13)

## Instructions (SKILL.md)

Use this skill when a brand or creator wants the "how did they get that giant shoe on the billboard" clip without a stunt crew: a fake out-of-home (FOOH) moment where the user's own product, giant, is in a city street. It is two 5 second shots filmed like a phone clip: a wide establishing shot that shows the street and the stunt starting, then a low-angle hero moment as the product towers overhead. The product's real shape is locked by keyframes made from the user's photo and approved before any video is paid for, the finals have street sound, and a small "AI-made" tag is drawn over the whole clip. The clip is openly a staged CGI stunt, never passed off as real.

### What you need

- **One photo of the product** the user owns or has rights to show, with no people in it (crop them out first). Clean and sharp, the whole product in frame, no other brand's items. If the photo is blurred or half in frame, ask for a better one before spending anything.
- **A stunt** from the menu in `references/stunt-menu.md`: `billboard-burst` (out of a billboard), `hill-roll` (down a hill street), `rooftop-park` (parked on a rooftop), `tram-wrap` (wrapping a tram) or `harbour-rise` (rising from a harbour). Choose it from the product: round and boxy things roll, long things wrap, anything can park. No default: ask, or suggest one.
- **A type of street**, from the menu (`downtown`, `old-town`, `hill-street`, `tram-street`, `harbour-front`) or in the user's own words ("a wide boulevard with plane trees"). Never a named place or a famous landmark: the clip must not read as a real event somewhere real.
- **Light:** `golden-hour` (default), `overcast-day`, `bright-morning` or `blue-hour`.
- **Logo:** left plain by default, because models garble letters. `--keep-logo yes` copies it from the photo; then check every letter, and re-roll or drop it if any letter is wrong.
- **Aspect** 9:16, **length** 8 to 10 s (two clips of up to 5 s, the hero trimmed if the model starts to invent writing), **resolution** 720p.
- **Limits: people.** Pedestrians are tiny, far away, seen from behind or at the frame edge, calm, with no face readable. Never a real person, celebrity or public figure, and never from a photo of a person. Seedance on fal refuses a start frame with a clearly visible realistic face, so the 480p drafts and the Seedance finals need the same rule; Kling accepts faces but the format does not need them.
- **Limits: honesty and safety.** No real third-party brands, billboards, shopfronts or vehicle badges. No famous landmark or named city. No damage, panic or emergency (no broken billboard, debris, smoke, fire, sirens, running crowds). No news, live-TV, CCTV or dashcam look. The AI-made tag is on by default; `--tag off` only if the user asks, and then they must still post with the platform's AI label and a caption that says it is CGI (`references/safety-and-honesty.md`).
- Optional: music the user owns or has licensed, added in the app; a caption written from `assets/caption-template.md`.

### What you make

1. `stunt.json`: the plan (locks, prompts, sounds, trims, tag).
2. Two approved keyframes (JPEG, 9:16, 2K): the wide frame and the low-angle hero frame.
3. Two optional 480p drafts and two 5 s Kling 3 pro finals with sound.
4. The stunt clip: H.264 MP4 with AAC sound, 720 x 1280, 24 fps, 8 to 9.7 s, one hard cut, the AI-made tag on every frame. Optional 4:5 crop and a silent web preview under 8 MB.
5. A caption that says it is CGI, from `assets/caption-template.md`.

### How to run it

**Budget: about $2.70 for a stunt on fal if every shot passes first time (keyframes $0.30, drafts $0.72, finals $1.68). Expect re-rolls: the sample run, without drafts, cost $4.95 (four keyframes, four finals). Quote $2.70 and warn the user it can reach about $5. Seedance 2.5 finals instead of Kling make it about $5.75 before re-rolls. A Kling 3 pro final took 4 to 8 minutes to come back in the sample run.**

Before the first paid call, run `node scripts/build-requests.mjs --plan stunt.json --stage estimate`, quote the total it prints, and get one explicit yes on it. Then, at every paid step, tell the user the model, the number of calls and the cost, and ask again if the spend will go past the quoted total. To send a request, POST the printed JSON `input` to the fal model id in that request with the user's own FAL_KEY (`scripts/fal-run.mjs --request <file> --out <dir> --approved` does this and embeds local photos as data URIs; the Skill never stores keys), or paste the prompt and images into any Nano-Banana-class image editor and Kling- or Seedance-class image-to-video tool that makes sound.

1. **Brief (free).** Look at the photo. Confirm the product, the stunt (does it suit the product?), the street type, the light and the logo decision. Refuse a product that is a weapon, tobacco or vape, alcohol, a medicine or anything an ad rule restricts, a photo with a person in it, and a street that names a real place or landmark (`references/safety-and-honesty.md`). Tell the user up front that the clip carries an AI-made tag and a CGI caption.
2. **Plan (free).** `node scripts/build-plan.mjs --out stunt.json --product "our running shoe" --photo shoe.jpg --stunt billboard-burst [--street downtown] [--light golden-hour] [--colour "sage green"] [--keep-logo yes]`. It copies the stunt's two shots, prompts and sounds from `assets/stunts.json`. Write `locks.product`, the only `FILL:` field: the product's shape, proportions, parts and colours from the photo, with no logo or brand name (`references/prompt-patterns.md`). Read the prompts against `references/shot-grammar.md` and describe the two shots to the user in plain words: what the street shows, what moves, what we hear.
3. **Spend gate (free).** `node scripts/build-requests.mjs --plan stunt.json --stage estimate`. Quote the printed total and wait for one explicit yes.
4. **Wide keyframe (about $0.15).** `node scripts/build-requests.mjs --plan stunt.json --stage keyframe --shot 1 --out requests --inline` (`fal-ai/nano-banana-pro/edit`, the photo as reference, `aspect_ratio` "9:16", `resolution` "2K"). Open it beside the photo: the product's shape and colours match, it is already at the key moment of the stunt (the poised start frame), it reads as giant next to the buildings and tiny pedestrians, the light and contact shadows are true, the street is generic with no readable text, logo or badge, nobody is large or readable, nothing is damaged, and the product sits in the middle 60% of the frame. Re-roll by changing one thing (max 3 attempts), then put the approved path in shot 1's `keyframe`.
5. **Hero keyframe (about $0.15).** `--stage keyframe --shot 2` (reference 1 the approved wide frame, reference 2 the photo). Compare it with the wide frame: the same street, building, set piece, light and product; low and looking up; the product towering with wide-lens perspective. Put the approved path in shot 2's `keyframe` and show the user both stills before any video.
6. **Drafts (about $0.36 each).** `--stage draft` (Seedance 2.0 mini, 480p, 5 s, silent). Make a contact sheet per clip (`ffmpeg -i draft.mp4 -vf "fps=2,scale=180:-2,tile=8x2" -frames:v 1 sheet.jpg`) and look: the move reads, the product keeps its shape and size, nothing breaks or spawns, people stay tiny. A different model from the final, so judge the idea, not the physics or sound. (The sample run skipped the drafts to keep the budget for Kling re-rolls, so this step is untested there.) If Seedance refuses the frame (a face), shrink the nearest person and regenerate the keyframe, or skip the draft.
7. **Finals (about $0.84 each).** `--stage final` prints the Kling 3 pro request: `fal-ai/kling-video/v3/pro/image-to-video`, `start_image_url`, `duration` "5", `generate_audio` true and a negative prompt (it returned 1072 x 1928 in the sample run; assembly scales it to 720 x 1280). Kling is the default because it keeps shapes better than Veo and accepts people in frame. `--model seedance` gives Seedance 2.5 720p (about $2.37 each, premium). Contact-sheet each clip at 2 fps and listen. Note clean in and out points (skip the first 0.1 s; stop before the product changes size, the billboard cracks or grows writing, or a second product appears; Kling tends to add writing to a blank billboard after about 3 s) and write `clip`, `in` and `out` per shot in stunt.json. Fail path per shot: max 3 attempts, one change each (move wording, sound line, then the model), say which; then deliver the best and say what still differs. If shot 2 will not hold, trim it earlier or deliver shot 1 alone (`--only 1`) and say the clip is shorter.
8. **Assemble (free).** `node scripts/assemble-stunt.mjs --plan stunt.json --out <product>-stunt.mp4`. One hard cut, the clips' sound joined with a 0.08 s crossfade, a light `clean` grade, and the "AI-made" tag drawn over every frame (needs sharp; `npm i sharp` in the folder you run from). Web preview: `--audio none --max-mb 7.5`. Feed version: `--crop 4:5`. `--tag-pos top-right` moves the tag; `--tag off` only if the user asked. If the sound is too quiet or hot, `--gain <dB>`.
9. **Check (free).** `node scripts/check-stunt.mjs --file <product>-stunt.mp4 --edl <product>-stunt.mp4.edl.json --resolution 720p`, then open the contact sheet, watch the clip twice with sound and go through `references/output-spec.md`. A cloud clearing the sun or a passing tram can trip the flash check: look at the moment, and if it is real light, re-run with `--ignore <seconds>`.
10. **Deliver.** The clip, the two keyframes, stunt.json and the caption from `assets/caption-template.md`, with the product name filled in. Say which shots were re-rolled and anything still imperfect. Tell the user to switch on the platform's AI-generated label when they post (TikTok, Instagram and YouTube all ask for it on realistic AI video), to keep the CGI line in the first line of the caption, never to present it as a real installation, and that the sound is model-made, so they should listen before posting.
11. **Clean up.** Delete the user's product photo from the working folders unless they want it kept, and any public copy. fal keeps request inputs (including inline photos) and outputs on its side for a period: tell the user they can delete them from their fal dashboard or ask fal to, and to check fal's current retention policy.

### Files

- `references/stunt-menu.md`: the five stunts, which products suit each, the streets, the light, the set piece colour and safety notes per stunt.
- `references/shot-grammar.md`: the two shots, keeping the product locked, scale cues, people, phone-clip behaviour, safe zones, sound, trims.
- `references/prompt-patterns.md`: locks, keyframe and motion prompt anatomy, failure modes and fixes, drafts and finals, the fail path.
- `references/safety-and-honesty.md`: the AI-made tag, the CGI caption, what the clip must not show, Seedance and faces.
- `references/output-spec.md`: deliverables, sizes, costs, sound, the pass/fail checklist and the posting note.
- `assets/prompt-template.md`: the keyframe and motion prompts with placeholders.
- `assets/stunts.json`: the stunt, street and light menus with their prompts and sounds.
- `assets/generation-input.json`: model ids, sizes, defaults and costs.
- `assets/caption-template.md`: a caption that says it is CGI, and the rules for it.
- `assets/example-input.jpg`: a synthetic, unbranded product photo (a white running shoe) to try the Skill with.
- `scripts/build-plan.mjs`: writes stunt.json from the product, photo, stunt and street. Refuses landmarks, named cities and people. Calls no API.
- `scripts/build-requests.mjs`: the cost estimate, and the keyframe, draft and final requests. Calls no API.
- `scripts/fal-run.mjs`: sends one request with the user's FAL_KEY after `--approved`, waits, downloads (never overwrites without `--force`).
- `scripts/assemble-stunt.mjs`: trims, one hard cut, audio crossfade, light grade, crop, the AI-made tag (sharp), H.264 and AAC with +faststart, edit list.
- `scripts/check-stunt.mjs`: technical checks, the tag, sound in every shot, black frames and flashes, contact sheet.

## Related

- [Cinematic brand film](https://meleon.ai/skills/cinematic-brand-film.md): Turn a one-line brand story into a graded 30 to 45 second film with a title card, the same cast in every shot.
- [Destination teaser](https://meleon.ai/skills/destination-teaser.md): Turn a place, a season and a mood into a graded 15 to 20 second vertical teaser, text-free and ready for captions.
- [ASMR cut reel](https://meleon.ai/skills/asmr-cut-reel.md): Remake an object or your product in glass, jelly, soap or chocolate and slice it in a looping ASMR reel of about 15 seconds with crisp cutting sound.
- [All Meleon Skills](https://meleon.ai/skills/) (index as JSON: https://meleon.ai/skills.json; site guide: https://meleon.ai/llms.txt)
